Just reposting the feedback I left you on your blog:
Howdy Kapil,
Thanks for bringing this up again on vpsBoard. I had forgotten to respond.
I believe you are reading your table incorrectly. Your table seems to be a representation of the spread of the times requests took, i.e. this data:
Percentage of the requests served within a certain time (ms)
50% 4414
66% 4488
75% 4543
80% 4578
90% 8705
95% 8948
98% 9054
99% 9118
100% 9252 (longest request)
It has no relation to increasing concurrency, as concurrency was kept the same throughout your test (at 100 connections). I do not think that the x-axis is time, but rather that these requests were ordered by the length of time they took. That is why, on the table, requests only get slower as you move along the x-axis.
What the table does show is that OpenLiteSpeed's slowest requests were slower than the other setups' slowest requests. (This may be where your 15% difference number is from.) However, the table also shows that the majority of requests served by OpenLiteSpeed (about 75%?) were served faster than the other setups. Also, on average, OpenLiteSpeed served requests marginally faster than both other setups, as noted by the average requests per second and median connection times data.
I'm not sure if I'll be able to find someone with time to look at what can be tweaked on your settings, though that might be interesting. I'll definitely see if there's someone here interested in taking that on.
Cheers,
Michael